Sources: Nvidia chips remain the norm for Chinese AI labs training LLMs, as switching from its CUDA platform to Huawei's CANN requires major code rewriting (South China Morning Post)

Observations from multiple sources show that Chinese AI labs have not shifted away from Nvidia's processors for LLM training. The continued use of Nvidia hardware suggests that performance, software maturity, or ecosystem factors are still preferred. No public announcements indicate a widespread migration to Huawei's CANN. Hence the hardware landscape for Chinese LLM development remains unchanged.

Adopting Huawei's CANN in place of Nvidia's CUDA would require substantial modifications to existing codebases. The rewriting effort is described as major, implying significant engineering time and cost. This cost deters labs from pursuing the switch despite potential strategic or supply-chain motivations. As a result, the barrier to entry for Huawei's platform remains high.

If labs attempt to run their current LLM code on Huawei's CANN without modification, the software will not execute correctly. The incompatibility stems from differences in the programming models and libraries between CUDA and CANN. Therefore, the switch only becomes viable after a comprehensive code porting effort. Until that work is completed, Nvidia's chips remain the practical choice.